Search Results

Collaborative filtering has been successfully applied for predicting a person's preference on an item, by aggregating community preference on the item. Typically, collaborative filtering systems are based on based on quantitative preference modeling, which requires users to express their preferences in absolute numerical ratings. However, quantitative user ratings are known to be biased and inconsistent and also significantly more burdensome to the user than the alternative qualitative preference modeling, requiring only to specify relative preferences between the item pair. More specifically, we identify three main components of collaborative filtering-- preference representation, aggregation, and similarity computation, and view each component from a qualitative perspective. From this perspective, we build a framework, which collects only qualitative feedbacks from users. Our rating-oblivious framework was empirically validated to have comparable prediction accuracies to an (impractical) upper bound accuracy obtained by collaborative filtering system using ratings.